Specialize the curvfit / surface fit rate models
The rate models for the curvfit and surffit models are
specialized by block size, by dividing the block sizes
into 4 categories. The prediction error for rate is
reduced substantially in the process.
Note there is much less improvement in distortion by
the same process and so a generic model for all block
sizes are still used for distortion. However the
distortion model for surffit is also updated by sampling
a little more coarsely than before.
lowres 20 frames end-usage=q: +0.023% av PSNR
midres 20 frames end-usage=q: -0.053% av PSNR
Upto 3% speed-up but bot consistently across all videos.
STATS_CHANGED
Change-Id: I1ef7335b2e40308e11790dd9fc13d062d4fc81c1
diff --git a/av1/encoder/rd.h b/av1/encoder/rd.h
index b2360ee..e29a1d5 100644
--- a/av1/encoder/rd.h
+++ b/av1/encoder/rd.h
@@ -656,9 +656,10 @@
void av1_model_rd_from_var_lapndz(int64_t var, unsigned int n,
unsigned int qstep, int *rate, int64_t *dist);
-void av1_model_rd_curvfit(double xqr, double *rate_f, double *distbysse_f);
-void av1_model_rd_surffit(double xm, double yl, double *rate_f,
+void av1_model_rd_curvfit(BLOCK_SIZE bsize, double xqr, double *rate_f,
double *distbysse_f);
+void av1_model_rd_surffit(BLOCK_SIZE bsize, double xm, double yl,
+ double *rate_f, double *distbysse_f);
int av1_get_switchable_rate(const AV1_COMMON *const cm, MACROBLOCK *x,
const MACROBLOCKD *xd);
